[Bench] Group benchmark rows by workload label and keep manifest order

[lcy-seso]

Problems

• An op's rows were sorted by case id, so a size sweep read out of order: FusedMoeSharedExpertFwd listed t1, t128, t2048, t32, t4096, t512, t64.
• Ops with several tensor templates (BmmFp8Fwd, GroupedQueryAttentionPrefillPagedWithKVCacheFwd, FusedTopKFwd, GemmFwd) numbered the table in one order and the key in another, so W4 sat above W2 in the key.
• Every dtype of a label was its own W row, so a reader had to look up the code in the key to find which label and dtype a number belonged to.
• A key row gave only symbol values (B=[2048], normalized_shape=[4096]), so reading a shape meant substituting them into [*B, *normalized_shape] by hand.
• A cluster with one label moved all its values above the label, so they read as belonging to the label before it (LayerNormFwd llama-13b-prefill, dit-xl-2).
• A wrapped key row began with a middot, and tensors and scalars broke wherever the width ran out; labels over 32 characters wrapped in the label column.
• Checked against TileOPs main (fc6f03c59), the hand-written pages named stale labels and roofline formulas, a kernel __call__ main replaced with forward, wrong test-mark roles, a device-keyed memo for every kernel, a bounded backend memo table, copies the timer supposedly never sees, and H200 bandwidth 4.07 TB/s where the profile now gives 4.50.
• Adding a Backend described an older tileops-backend-example: kernels.py, a pending.py under a misspelt key, _detect, four error paths, and two requires_cuda_runtime tests with 22 or 20 passing.

Changes

• Rows keep the snapshot order, which is the manifest order, and the table follows the key cluster by cluster.
• Each manifest label is one row group: the label once, spanning its dtype rows, then a dtype column (fp8e4m3/bf16 for a case with two dtype indices). The W codes are gone.
• The key lists each label once with the values that vary and the dtypes it ran at; a dtype parameter already in the case id (out_dtype, cache_dtype) is not repeated there.
• A tensor's dtype mark is relative to the dtype the row shows, and one label whose dtypes differ only in that mark stays one group.
• A row with no manifest entry (a legacy snapshot, an undeclared op, no TileOPs checkout) takes its id with the trailing dtype names split off as its label.
• Long labels wrap at hyphens, capped at 18em in the table; on a narrow screen the key puts each label on its own line.
• Each key row prints its shapes with the symbols substituted (x: [2048, 4096]); the template stays once above the labels, and the symbol values are no longer listed.
• Values move above the labels only when two or more labels share them.
• A line of key entries clips the middot of the entry that opens it, scalars wrap as one part after the tensors, and a cluster with a label over 32 characters sets each label on its own line.
• Writing a Spec, Adding an Op, torch.compile, Adding a Backend, Timing, the home page, the API index and the memory-bound guide (en and zh) now state what main does; the roofline figure is redrawn to 4.50 TB/s. Adding a Backend follows the current example: target.py, one module per op under ops/, a BUILDERS table, three error paths, 24 tests passing with and without a visible GPU. Worked examples the manual still lacks (type families, ADTs, generators with requires, composition, multi-case dtype_cases) are left to a separate PR.
• The test fixture gains a second dtype row, and the goldens are refreshed for the new layout and the substituted shapes. Test node delta: 0 (10 before, 10 after).

Before and after, FusedMoeSharedExpertFwd on the 2026-09-27 snapshot:

before                                   after
| W1 | 1.16× | 0.2135 |   deepseek-v3-t1     | bf16 | 1.16× | 0.2135 |
| W2 | 1.04× | 3.0098 |   deepseek-v3-t32    | bf16 | 1.00× | 2.0624 |
| W3 | 1.10× | 5.7252 |                      | f16  | 1.00× | 2.0636 |
| W4 | 1.00× | 2.0624 |   deepseek-v3-t64    | bf16 | 0.98× | 2.5399 |
| W5 | 1.00× | 2.0636 |   deepseek-v3-t128   | bf16 | 1.04× | 3.0098 |
| W6 | 1.15× | 8.1639 |   deepseek-v3-t512   | bf16 | 1.03× | 3.9286 |
| W7 | 1.03× | 3.9286 |   deepseek-v3-t2048  | bf16 | 1.10× | 5.7252 |
| W8 | 0.98× | 2.5399 |   deepseek-v3-t4096  | bf16 | 1.15× | 8.1639 |
(W1-W8 = t1, t128, t2048, t32 bf16, t32 f16, t4096, t512, t64)

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.